Self-supervised Monocular Depth and Pose Estimation for Endoscopy with Latent Priors
Researchers have developed a new self-supervised framework for estimating depth and pose in endoscopic videos. This method utilizes a Generative Latent Bank trained on natural images to improve depth prediction realism and robustness. Additionally, it reframes pose estimation within a Variational Autoencoder to stabilize predictions and enhance accuracy. Evaluations on SimCol and EndoSLAM datasets show this approach outperforms existing self-supervised methods for endoscopic applications. AI
IMPACT Enhances AI's ability to provide precise 3D mapping for medical diagnostics and procedures.